object: [shape:spherical color:blue texture:smooth]
The object of the task is to associate the object with the label
'ball', generalizing the concept (e.g., color is not related to the
final concept of ball). Several constraints on the design of the
system were identified, including placing the model within a unified theory of cognition and using an incremental learning method like Soar's
basic chunking mechanism.The problem space for concept acquisition consisted of operators to recognize descriptions and propose labels (concepts). Search in the problem space consisted of the piece-wise removal of attribute-value pairs in the description until a concept was found. During training, in which the concept was included with a description, chunks were built that progressively removed these attribute-value pairs to generalize the concept. This was a form of abstraction and was generated dynamically, in response to impasses. After training, concepts were presented and recognitional chunks fired on the presentation of the description to identify the concept. One limitation to this approach is that it requires the specification of all possible concepts before training begins. However, it does show that Soar can take on inductive tasks and, in this case, perform similarly to systems (e.g., ID3 and COBWEB) that were designed specifically for concept acquisition tasks. Additionally, many of the empirical results from experiments with this system were consistent with psychological data, strengthening Soar's claim as a unified theory of cognition.
Go to a discussion of this capability for multiple architectures.